Follow the crowd : Social information and crowdfunding donations in a large field experiment

esearch shows that donors tend to mirror the donation amounts of others. This behavior, adjusting the donation amount to the donation behavior of others, is described as the social information effect (Shang & Croson, 2009). A recent systematic literature review (van Teunenbroek, Bekkers & Beersma, 2020) shows that donors informed about the donation amount of previous donors tend to donate higher amounts. While overall previous research tends to find positive effects of social information, its effect depends on several contextual factors. First, prospective donors who find the amount donated by others to be excessively high may refrain from giving altogether (Croson & Shang, 2013), while an amount that is too low may lower donation amounts (Croson & Shang, 2008; Meyer & Yang, 2015). In addition, the published research shows that whether social information encourages donation behavior depends on how donors interpret it, and they may do so in different ways. Several studies suggest that social information provides donors with a norm that guides their donation behavior (Croson, Handy & Shang, 2009; Smith, Windmeijer & Wright, 2015); such that donors think that previous donations indicate a standard for what is appropriate. We provide further evidence on the effects of social information on donation behavior in a largescale field experiment. Our paper contributes to the literature in two ways. First, we examined the effects of stating donation amounts of earlier donors in a R Abstract: Purposely guiding human decision making with a discrete suggestion, ‘nudging’, is increasingly popular. One particularly promising nudge is to provide decision makers with information about the decisions of others, also referred to as social information. Social information is often applied in fundraising campaigns to increase individual donations. A discrete suggestion such as the donation amount of others can result in donors donating similar amounts. We examined effects of social information in a relatively new context, namely crowdfunding. Crowdfunding is a new online fundraising tool. Our study, based on a large natural field experiment (n = 24,070), tests to what extent social information affects online donation behavior and how its effects vary throughout the duration of a campaign. We show that social information increases the individual donation amount by 17%, which is close to the average of 14% found in previous studies. However, social information did not attract more donors: the participation rate was not affected. Our study is the first to pinpoint the stage of the funding campaign at which the effect of social information is most pronounced. We found that social information is most effective in increasing donations at the beginning of crowdfunding campaigns. All materials for this article are available at https://osf.io/epuj6/.

context in which it has hardly been tested before, namely crowdfunding. Crowdfunding is a new online fundraising tool that builds on small donations from a large (and mostly unknown) crowd (Mollick, 2014). Our research question is: "what is the influence of social information on online donation behavior through a crowdfunding platform?" Second, we provide evidence on the optimal timing of social information. Crowdfunding campaigns are launched online at a platform for a specific duration. Our study is the first to pinpoint the stage of the funding campaign at which the effect of social information is most pronounced. Was social information more effective when the campaign just launched, or towards the end? An answer to this question adds further evidence to the stock of knowledge on contexts in which 'nudges' such as the provision of social information work.
The implications are important for practitioners since there is a need to understand possible stimulants for donating to crowdfunding projects (Zvilichovsky, Danziger, & Steinhart, 2018), as many crowdfunding projects fail to assemble enough funding. For instance, between 2014 and 2018 about twothird of the projects on one of the most popular and successful crowdfunding platforms, Kickstarter, failed to assemble the target amount (The Crowdfunding Center, 2018). Second, and more broadly, crowdfunding is a private source of income that may replace government funding for the arts. After a large cut in government funding for the arts in the Netherlands, the sector struggled financially (Blankers, Goudriaan, de Groot, Everhardt, Friperson, & Mazzola, 2012). To reduce their dependence on government funding, arts organizations are forced to acquire income from alternative sources such as donations (Algemene Rekenkamer, 2015). Our findings could be used to increase the effectiveness of crowdfunding campaigns.
To test the effect of social information, we conducted a large-scale field experiment among all visitors (n = 24,070) on a Dutch crowdfunding platform. The experiment tested the effects of social information on donation behavior. In the treatment condition, the message "Did you know that on average donors on Voordekunst donate 82 euros?" was displayed next to the project information. This was the actual average amount donated by donors on the platform of the preceding six months. In the 'base' condition, there was no visible average donation amount (see appendix A for stimulus materials).

The Effects of Social Information in Previous Studies
Donors adjust their charitable behavior according to social information. When individuals are presented with information on the donation amount of previous donors their donation amount increases (for example Alpizar, Carlsson, & Johansson-Stenman, 2008a;Bekkers, 2012;Edwards & List, 2013;Martin & Randal, 2008;Shang, Croson & Reed, 2012;van Teunenbroek, 2016;van Teunenbroek et al, 2020).
The prevailing explanation why social information works is that information about the decision of others creates a social norm (Bøg, Harmgart, Huck, & Jeffers, 2012;Croson et al., 2009;Croson & Shang, 2008, 2013Edwards & List, 2013;Meyer & Yang, 2015;Murphy et al, 2015;van Teunenbroek et al., 2020;Sasaki, 2018;Smith et al., 2015). The average donation amount provides a cue about what is typically done by others. A classic premise in social psychology is that humans have a strong desire to follow social norms and mirror the behavior of others (Bernheim, 1994;Festinger, 1954). According to social comparison theory, humans evaluate themselves in comparison with others, for instance to reduce uncertainty (Festinger, 1954). Thus, human decisionmaking is influenced by social norms and people mimic the behavior of others. When the default is not to give or to give less than the norm, social information increases donations. Accordingly, we propose that: Hypothesis 1: Social information increases the amount donated.

How the Effect of Social Information Varies with the Project Funding Stage
A crowdfunding campaign runs for a predefined number of days, during which the target amount must be assembled to be considered successful. Taking the project funding stage into account is important, since the effects of social information could vary with the fundraising stage. An earlier study found that social information in the form of fake (i.e. created to mislead consumers) Facebook Likes differently affected the number of donors to crowdfunding campaigns depending on the fundraising stage: initially there was a positive effect, followed by a negative effect over time (Wessel, Thies & Benlian, 2016). While the social information observed by Wessel et al. (2016) differs from mentioning the donation amount of others, it does show the importance of the funding stage. We assume that the donors in the middle of the campaign are especially sensitive to our form of social information. The assumption is based on the idea that social information provides a quality signal to donors (van Teunenbroek et al., 2020;Vesterlund, 2003), as argued in theories on philanthropy from communication science, behavioral economics, and social psychology. Initial stage. At the initial stage, a project is seen as new and innovative which attracts individuals motivated by contributing to new ideas (Rogers, 1995). These donors are not motived by the example of others. Instead, they want to lead and they want to be first: they are the 'early adopters'. Rogers argues that innovations spread through communication among peers. In our case, the number of donors and the amount assembled per day is expected to be relatively high in the beginning of a campaign, see Figure 1. A similar donor type is described in the philanthropic literature: the impact donor cares about the impact of her donation on the provision of a public good. Impact donors want to be pivotal, and care about the difference that their donation makes (Duncan, 2004). The impact motive explains why some donors prefer to fund a specific part of a project rather than the entire project. The impact donor finds a project less attractive if more donors have contributed, since the contributions of other donors reduce the impact of her additional donation. This implies

Figure 1 Expected Distribution of the Social Information Effect and Amount Donated
that for impact donors, social information in the initial stage will have no or even a negative effect. In addition, in the beginning of the campaign the crowdfunding project is mostly supported by family and friends who have strong ties to the creator. They donate because they have a close connection with the creator (Borst, Moser, & Ferguson, 2018) and not so much because they want to make a difference or support a high-quality project.
Middle stage. After the group of innovative enthusiasts and strong ties from the creators' network is exhausted, a different type of donor must step in. The crowdfunding literature describes that at this point, in the middle of the campaign, the number of donations per day decreases (Kuppuswamy & Bayus, 2015). This is the stage where social information could have the greatest impact, because identification with the creator is relatively weak. In the middle stage of crowdfunding projects donors are much less likely to have a social connection with the creator (Borst et al., 2018;de Witt, 2012). When identification is weak, individuals mainly base their donation decision on the perceived quality of the project and they wonder whether the project is worthwhile (Fishbach, Henderson, Koo, 2011). We expect that social information could be particularly effective at this stage, since social information signals that others value the project enough to support it (Vesterlund, 2003). Knowing that others have already donated, new donors can assume that other donors have checked it out and rest assured that the project is worth supporting.
End stage. Throughout the fundraising period, the distance to the target amount decreases with each donation. As the total number of donors reaches a certain threshold it becomes increasingly attractive for individuals who want to interact in a successful social setting (Oliver, Marwell, & Teixeira, 1985). In other words, by donating at this stage, donors become part of the critical mass that reaches the target amount (Markus, 1987). As the number of previous donors to a project increases, so do the odds of a new donor making a donation (Oliver, & Marwell, 1988): the closer a crowdfunding campaign comes to the target amount the higher the participation rate (Kuppuswamy & Bayus, 2017;Zvilichovsky et al., 2018).
Our project funding stage hypothesis proposes that: Hypothesis 2: Social information is less effective in the beginning as well as towards the end of the campaign, and most effective in the middle of the campaign.

Study Context
The data we analyzed was collected at Voordekunst, the largest crowdfunding platform in the Netherlands for cultural and arts projects, including for example dance, photography, music, theatre, movies and visual arts productions. More information about the platform can be found in Appendix B. Below we report how we determined our sample size, all data exclusions, all manipulations, and all measures in the study.
We measure two dependent variables: 1) individual donation amount; and the 2) number of donors. Visitors to the website were randomly distributed (50:50) over two conditions using browser cookies. Cookies ensure that participants using a specific desktop end up in the same condition each time they visit the website, regardless of the project(s) they view or what entry to the website (e.g., through social media, email or a direct URL visit) they used. Our treatment (Appendix A) is the addition of social information to all projects advertised on the platform with the following sentence: "Did you know that the average donation amount at Voordekunst is €82?". Convinced by Hertwig & Ortmann (2001), we used the principle of no deception, and showed website visitors in the treatment group the actual average amount donated by donors on the platform in the preceding six months.
Because it was not possible to include a manipulation check after donors finalized the payment procedure on the platform, we do not know if the participants paid attention to the social information given in the treatment condition. However, the only difference between the control and the treatment group is the provision of social information. Therefore, it is reasonable to assume that a difference inbehavior between the two conditions is a result of the manipulation in the treatment condition.

Figure 2 Histogram of the amounts donated per condition.
Note: A Chi square test indicates that social information did not affect the distribution of amounts over all categories (significance indicated by a * at p<.05. The experiment was only conducted among desktop users, ignoring other devices such as mobile phones and tablets. While we wanted to include all types of devices, due to technical limitations this was not possible. Data provided by the platform revealed that about two thirds of donors use a desktop (64%).
Crowdfunding platforms host many projects and each project receives donations from many donors. The data assembled give us a unique opportunity to test for these project effects. We exploit the natural variation in crowdfunding at two levels: project funding time (i.e., the number of days since the project launched, at the time of the individual donation), and projects. Because the data were completely anonymized, we do not have information identifying website visitors or donors.

Study Procedure
On the crowdfunding platform, each project has a separate webpage. If a participant decides to donate on a project page, the donor is sent from the 'project page' to the 'donation page'. In the treatment condition, the average donation amount is mentioned on both pages at the same place and with similar framing (see appendix A).
On the donation page, the donor specifies the amount she wants to donate, and whether she wants to receive a reward if the donation is high enough to receive a reward. The participant is then guided towards the 'payment page'. On this page, no social information was mentioned.

Participants
The experiment was conducted among 24,070 unique website visitors of the Voordekunst platform in September 2016. 11,973 website visitors were assigned to the treatment condition and 12,097 in the control condition. 2,657 website visitors (11.0%) donated, 1,374 in the control group (51.7%) and 1,283 in the treatment condition (48.3%).

Data Description
The data consists of visitors and donors (see Table  1). Website visitors are all individuals who visited the platform using a desktop within our period. Donors are all individuals who donated using a desktop in the time of our period. Each donation is made by a donor to a specific project. During the period of our experiment, donations were made to n = 119 different projects.
As the histogram in Figure 2 shows, the amounts donated are not normally distributed, giving rise to the possibility that a few very large donations greatly affect the total amount donated. We used a single-construct technique of standard deviation analysis (Aguinis, Gottfredson, & Joo, 2013), and considered a data point as an outlier if it was more than three standard deviations from the mean. While most donations hover around 25 euros for both the control and treatment condition, there were a few exceptionally large donations (1.39%, n = 19), more than three standard deviations above the average. Therefore, we focus on median amounts, and use a Mann-Whitney U Test unless otherwise noted. See Appendix E for a robustness analysis where we explore the data with a model including the natural logs of the amounts donated.

Social Information Increases Amounts Donated
We hypothesized that social information increased individual donation amounts. This hypothesis was supported by the data (U = 836375.50, n = 2,657, Z = -2.289, p = .022, r=.04). The median donation in the treatment group (€35) was 17% higher than in the control group (€30).
A regression analysis of the natural log of the amounts donated with project fixed effects confirmed that social information increased amounts donated (b = 3.68, n = 2,657, p = .018). Social information did not stimulate more people to give. There was no difference in the number of website visitors who donated between the treatment and control condition (X 2 = 1.50, n = 24,070, p = .221). The conversion rate for the control condition was 11.4% and 10.7% for the treatment condition.

Social Information Effects Throughout the Campaign
To test the project stage hypothesis, we compared the amounts donated in the treatment and control group in the beginning, middle, and end stages of crowdfunding projects (Table 2; Appendix D provides a more elaborate description).
The median amount donated was highest in the middle stage (see Table 2). While median amounts in the middle stage were 25% higher in the treatment than the control condition, the effect was not significant in this stage, U = 29327.00, n = 495, p = .421. The effect was only significant in the beginning stage (U = 278984.50, n = 1,544, p = .039), when median donation amounts were 20% higher in the treatment condition than in the control condition. Social information did not increase donations in the end stage, with median donations being slightly lower (8%) in the treatment condition than in the control. The effect of social information was not significant at this stage, U = 46898.50, n = 618, p = .714. Next, we tested whether the effect of social information on donation amounts was moderated by the project funding scale. Given the skewness of the data, we used the natural log of the donation amounts. A regression showed a negative but insignificant effect of the moderation effect (b = -.01, n = 2,657, p = .769). Figure 2 shows the distribution of the amounts donated per condition. The difference between the control and the treatment group was not as predicted. In the treatment condition, only one donation of exactly €82 was made, and fewer donations in the category of €81-90 were made in the treatment group (0.6%) than in the control group (0.9%). This result indicates that the amount we mentioned did not constitute a social norm that the participants followed. Instead, the treatment seems to have increased the number of donations that are clearly higher than the average donation of €82 we mentioned. In the treatment condition, 9.0% donated between €91 and €100 vs 7.6% in the control group, but not significantly so (see Figure  2). The treatment significantly lowered the proportion of donations up to €25 (45.1% in control vs 39.3% in treatment). Importantly, the treatment significantly increased amounts between €501 and €10-00 (1.7% vs 2.7%).

Discussion and Conclusion
Our analysis revealed support for the main hypothesis, that social information increases amounts donated. We found a modestly positive effect of about 17%, which is close to the 14% found in previous studies. However, social information did not attract more donors: the participation rate was unaffected. This is an important result because nudges can also backfire in the form of a lower likelihood of donations when donors consider them as coercive (Goswami & Urminsky, 2016). The social information we provided did not scare away donors who planned to give lower amounts.
In our project stage hypothesis, we predicted that the effect of social information would be strongest in the middle of the campaign, in which most donors consist of individuals searching for cues. While we found individuals in the middle of the campaign to give 25% higher amounts when presented with social information, this difference was not significant. We only found a significant effect (20% increase) of social information at the beginning of crowdfunding campaigns. Our assumption that individuals in the middle of the campaign are more uncertain about the quality of the project and therefore rely more heavily on social information than donors in the beginning stage may be incorrect. In addition, our assumption that friends and family are less strongly affected by social information may also be incorrect. We know from an earlier study of donations on the same platform with similar projects that donors in the beginning and end stage of a crowdfunding project mainly consist of family and friends, while donors in the middle of the project are mostly unknown to the creator (Borst et al., 2018). Our current results indicate that social information is most effective in the beginning of crowdfunding campaigns but continues to be effective in the middle and end stage. This pattern is unlikely to be a result of differences in the social ties with creators alone.
We explored whether the specific amount we mentioned (€82) would set a norm that others follow by donating the exact same amount as found by Sasaki (2019), van Teunenbroek (2016 and Bekkers (2012). We clearly did not find such a pattern. Perhaps the discrepancy is a result of the amount we suggested being a peculiar number (€82) rather than a round number such as €35 or ¥10,000. In any case, our results suggest that social information can affect giving even when it does not create conformity. Alternatively, it is possible that social information incr-eases the awareness of the need for donations and provides a signal of quality (van Teunenbroek et al., 2019). Because we could not include a manipulation check in the field experiment, future research is required to test these explanations.

Limitations
One limitation of the study is that we were unable to rule out contamination due to the field setting. First, participants in the treatment group may have shared the social information with participants in the control group. While we cannot rule out that contamination occurred, we do not have indications that participants were troubled by the manipulation and discussed it with others. Second, website visitors that disable or remove cookies after a browsing session, browse in incognito mode, or use different browsers may be exposed to different conditions. Data collected by the platform indicates that the proportion of recurring donors on the platform was low (13%) (Voordekunst, 2016a). While we have received no questions or remarks from website visitors about peculiarities in the design of the platform, it is possible that some users were exposed to different conditions. This may have weakened the effect of our manipulation, and the effect size we obtained may be an underestimation.
A second limitation of our experiment is that we have no information about the characteristics of individual donors. We considered adding a short survey to collect individual donor data. However, we opted not to do so because in all likelihood the response rate would be very low and selective. More research is needed to specify segments of donors for whom social information is particularly effective.

Project characteristics
At each project page, a visitor can see additional information about the project: (1) the amount donated thus far, (2) the number of donors thus far, and (3) the percentage of the target amount assembled thus far. We explored the robustness of the effect of social information with respect to these project characteristics in a regression analysis analyzing the natural log of donation amounts reported in Table  4. We provide estimates for two models: a. including all observations; b. excluding outliers. As in the main analyses, the results show positive coefficients for social information in both models. The significance level is sensitive to inclusion of outliers (p = .051 including all donations, p = .029 excluding outliers). The results also show that the total amount of previous donations is related to the amount donated by new donors, but the effect is miniscule. Model 3 shows a negative relationship between the amount donated and the total number of donations made to a project. However, the effect is small. The percentage of the target amount assembled is not related to the amount donated. In sum: both the previously donated amount and the number of donations are related to the amounts donated. However, the relationship with the amount donated is miniscule and the effect of social information holds while including these additional information aspects.